Sleeper agent (backdoored model)
A model that behaves normally until a specific trigger appears in its input, then follows a hidden instruction planted in its weights.
The term comes from Anthropic's 2024 'sleeper agents' paper and was made concrete by Microsoft's Lazarus research in November 2026, which implanted a shell command by editing fewer than ten dormant neurons without retraining. Sleeper behaviour passes standard evaluations because the trigger never appears in them; detection relies on comparing a checkpoint with its original weights or on scanners trained on known edit patterns, neither of which can prove a model clean.
Related terms
AgentAgent harnessClosed modelConfidential computingFine-tuning
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.